Testing for baseline differences in randomized controlled trials: an unhealthy research behavior that is hard to eradicate

Testing for baseline differences is often propagated because of the belief that it shows whether randomization was successful and it identifies real or important differences between treatment arms that should be accounted for in the statistical analyses. Especially the latter argument is flawed, because it ignores the fact that the prognostic strength of a variable is also important when the interest is in adjustment for confounding. In addition, including prognostic variables as covariates can increase the precision of the effect estimate. This means that choosing covariates based on significance tests for baseline differences might lead to omissions of important covariates and, less importantly, to inclusion of irrelevant covariates in the analysis. We used data from four supermarket trials on the effects of pricing strategies on fruit and vegetables purchases, to show that results from fully adjusted analyses sometimes do appreciably differ from results from analyses adjusted for significant baseline differences only. We propose to adjust for known or anticipated important prognostic variables. These could or should be pre-specified in trial protocols. Subsequently, authors should report results from the fully adjusted as well as crude analyses, especially for dichotomous and time to event data. Based on our arguments, which were illustrated by our findings, we propose that journals in and outside the field of nutrition behavior actively adopt the CONSORT 2010 statement on this topic by not publishing significance tests for baseline differences anymore.